§01 · Pipeline Overview
The muser-personal tool triages a Google Photos Takeout into three buckets — personal, in-between, and reference — using evidence-driven scoring over frozen embeddings. The diagrams below visualize each stage of the pipeline and the decision boundaries that emerge from the classification model.
Visualizations
Evidence → score → bucket flow with evidence reveal on hover
Base, decision, and combined sankey flows
Muser library + personal corpus merged via UMAP
Interactive point cloud of the embedding space
§02 · Read the Diagrams
Filter → Sort → Flow. The "Every image, down the stack" chart shows how 46,978 images flow through the pipeline. Width represents quantity; hover over the middle section to reveal the evidence signals (face presence, camera EXIF, document type, appearance) that drive the bucketing decision.
Evidence signals. Four signals feed the decision model: faces (5,759 images), camera EXIF metadata (2,508 images), document type (7,633 images), and visual appearance alone (31,078 images). Each signal carries its own distribution across buckets — personal images are face- and camera-heavy; reference images skew appearance-only.
Manifold structure. The combined embedding visualization shows 75k points (personal + main library + reference) merged into one cohesive space via UMAP with global-cohesion tuning. Personal images thread through the reference bulk as a red region, reflecting genuine visual overlap with reference imagery while maintaining distinct clustering.